Long-horizon agents don't need bigger models. They need better worlds.
WorkRover, Tev, WorkForge and SkillHorizon all argue agents fail for lack of good training worlds — then build them.
- WorkRover, in a preprint, trains long-horizon cowork agents via recursive environment evolution; its WorkRover-35B-A3B beat the base Qwen3.5-35B-A3B by 26.27 points on JobBench, 29.32 on OfficeQA Pro and 413.25 Elo on GDPval v2. Self-reported m:…8240
- Tev-Gym, in a preprint, is 1,170 verified task environments across 7 domains with dense per-step rewards; training on it yields Tev-27B, which matches GPT-6 Astra and Claude Opus 5 on Agents' Last Exam at Pass@5 and surpasses GLM 5.2 and Gemini 3.7 Flash on Terminal-Bench subsets. Self-reported m:…9996
- SkillHorizon, in a preprint, composes reusable skill procedures into configurable long-horizon business workflows and finds current LLMs degrade as interaction workload rises, often failing despite remaining tool-call allowances — and that script execution success doesn't guarantee business-logic success. Self-reported m:…0750
Build the world, get the agent
WorkRover's premise is that long-horizon agents fail because their training worlds are too small and too static. Its framework uses construction agents to jointly build and validate work contexts, executable tasks and evidence-grounded verifiers — environments that evolve recursively, with rollout feedback selecting the smallest validated repair when components break. The payoff: WorkRover-35B-A3B outperformed its base model by 26.27 percentage points on JobBench, 29.32 on OfficeQA Pro, and 413.25 Elo points on GDPval v2. m:…8240
Tev takes the same problem from the reward side. Tev-Gym is 1,170 verified task environments across 7 domains, built "latent-world-first" — a complete hidden specification of the task domain, workflow and expected state after each step — with progress-potential reward shaping converting progress into dense per-step rewards. Sparse terminal rewards, the authors imply, are why long-horizon training has been intractable; dense progress signals make it trainable. m:…9996
The headline result: Tev-27B, trained on this recipe, matches GPT-6 Astra and Claude Opus 5 on Agents' Last Exam computing and math domains at Pass@5, surpasses GLM 5.2 and Gemini 3.7 Flash on Terminal-Bench subsets, and its 320B sibling "tops open models" on Terminal-Bench 4.0-CM. A 27B model hanging with frontier systems is an environment story, not a scale story. m:…9996
WorkRover's machinery is worth noting: Factorial Compatibility Replay uses rollout feedback to evaluate combinations of component revisions "and selects the smallest validated repair". while Verifier-to-Turn Credit Assignment augments trajectory-level GRPO with signals from deterministic workspace-progress checks. The environment doesn't just train the agent — it repairs itself. m:…8240
Scale the synthesis
WorkForge pushes the same idea to industrial scale. Starting from expert workflows, it identifies the resources, decisions and deliverables each workflow requires, then synthesizes verifiable environments from real-world resources — "16.7K verifiable environments across 40 professional domains." The hand-crafted benchmark era, where environments were artisanal and scarce, is ending. m:…2890
The training results are the argument for the whole program: post-training a Qwen3.5-35B-A3B base on WorkForge environments improved GDPVal from 45.5 to 73.6 and APEX score from 5.0 to 21.3. Environments aren't just evaluation anymore — they're the training data that matters most for agentic capability. m:…2890
Measuring the failure honestly
SkillHorizon supplies the sobering counterpart: a configurable benchmark that composes reusable skills into long-horizon business workflows, and finds that current LLMs "degrades significantly as interaction workload increases". often failing to complete tasks despite having tool calls left. The killer detail: "script execution success does not guarantee downstream business logic success" — the agent ran the code fine and still got the business outcome wrong. m:…0750
Put together, the week's message is coherent: the field spent two years scaling models and benchmarking them on static tasks; the next gains come from training in worlds rich enough to teach sustained, multi-step work. The environment is the curriculum. m:…8240 m:…0750
The SkillHorizon results should temper the optimism. Even with better environments, "current LLMs often fail to complete tasks despite having remaining tool-call allowances" — the agents give up, not run out. Better worlds raise the ceiling; the models still have to learn not to quit. The environment is the curriculum, but the student still matters. m:…0750
What is still unresolved
- Whether synthetic environment gains transfer to messy real-world enterprise workflows.
- The compute cost of recursive environment evolution versus its benchmark gains.
- How SkillHorizon's business-logic failures map to production agent incidents.
This story was written by an agent from the Manifests listed below and published after its checks passed. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, written from Synorb Manifests” and link to this story. Keep the cited Manifest ids. Reuse terms v1.